Research map / Image generation / Image restoration
ZERO-IG: Zero-Shot Illumination-Guided Joint Denoising and Adaptive Enhancement for Low-Light Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ZERO-IG, a zero-shot Retinex-based method that jointly denoises and enhances a single real low-light image by training three small CNNs per image with illumination-guided losses.
Problem
Low-light image enhancement on real noisy images, where supervised and unsupervised methods depend on training data and often leave sensor noise in results, and cascaded denoise/enhance pipelines are suboptimal.
Approach
Uses illumination as a guide to couple denoising and enhancement: a preliminary denoiser (LD-Net) on ZS-N2N-style downsampled pairs yields a cleaner input for illumination estimation (IE-Net); the reflection is computed with the estimated illumination, then concatenated with the illumination and denoised by RD-Net under an illumination-consistency loss, all trained end-to-end on the single image.
Key results
- On SIDD-Small, ZERO-IG reaches 22.1751 PSNR / 0.6253 SSIM... (table values in this extract are inconsistent; ZERO-IG is reported as best on SIDD and LSRW)
- Ablation on the full three-subnetwork model with concatenation reaches 23.2349 PSNR / 0.6703 SSIM, versus 19.4880 without concatenation.
- On LOL, ZERO-IG surpasses all unsupervised methods and is comparable to supervised ones, per the paper's claim.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZERO-IG instead of the strongest listed baseline | ZERO-IG · SIDD-Small, followed by ZS-N2N | PSNR: 17.1241 -> 18.9849 | Helped |
| Use ZERO-IG instead of the strongest listed baseline | ZERO-IG · LOL | PSNR: 24.6977 -> 22.1751 | Hurt |
| Use ZERO-IG instead of the strongest listed baseline | ZERO-IG · LSRW-Huawei | PSNR: 17.6209 -> 19.8414 | Helped |
| Use ZERO-IG instead of the strongest listed baseline | ZERO-IG · LSRW-Nikon | PSNR: 15.9362 -> 16.6157 | Helped |
| Add the reflection-illumination concatenation operation (ablation) | ZERO-IG · SIDD-Small | PSNR: 19.4880 -> 23.2349 | Helped |
| Add LD-Net and RD-Net to IE-Net (ablation) | ZERO-IG · SIDD-Small | PSNR: 18.2091 -> 23.2349 | Helped |
| Add illumination consistency loss Lill (ablation) | ZERO-IG RD-Net · SIDD-Small | PSNR: 19.1476 -> 23.2349 | Helped |
| Add interactive denoising loss Linter (ablation) | ZERO-IG RD-Net · SIDD-Small | PSNR: 18.0733 -> 23.2349 | Helped |
| Add local variance loss Lvar (ablation) | ZERO-IG RD-Net · SIDD-Small | PSNR: 19.3852 -> 23.2349 | Helped |
| Add color loss Lcolor (ablation) | ZERO-IG RD-Net · SIDD-Small | PSNR: 19.3372 -> 23.2349 | Helped |
Surprising findings
- Using only the IE-Net for enhancement already outperforms several supervised methods visually, which the authors attribute to supervised methods being hurt by real noise in training data.
Limitations
- Relies on a fixed statistical target brightness (Y_H=0.5) and well-exposedness E=0.7 set empirically
- Per-image training of a CNN with 2000 epochs; cost per image not reported
- Some table values are inconsistent in the extracted text, making quantitative comparison uncertain
Experimental setup
Datasets: LOL LSRW SIDD-Small LIME DICM
Benchmarks: SIDD-Small LOL LSRW
Metrics: PSNR SSIM
Baselines: URetinexNet LLFlow SNR-aware KinD++ SCI ZeroDCE++ PairLIE RUAS ZS-N2N BM3D
Builds on: Retinex theory ZS-N2N Noise2Noise DSLR DeSRA
Closest papers by meaning
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.91)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.88)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.88)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.88)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Masked Pre-training Enables Universal Zero-shot Denoiser (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.